YXYang XiaoResearch portfolio
Yang XiaoForecasting & information
Research question

Would the forecast work in real time?

Economic statistics are revised after publication. A model can look better in hindsight if its backtest uses numbers no one knew at the time.

01 · Method

I reconstruct official release histories and rerun forecasts with three information rules: data known then, revised values with the same availability, and unrestricted latest values.

02 · Conclusion

Changing the data vintage changes which models look best.

Only the strict as-of rule avoids both future releases and future revisions. Even holding availability fixed, revised values change the GDP ranking of all six models.

Scope of the evidence. The fixed-mask comparison is a hindsight diagnostic, not a valid real-time forecast. Results describe this pilot and its final holdout; they do not establish a universally best model.

03 · Direct evidence

See what supports the finding.

Open interactive Space ↗

Hindsight enters through two doors

7,985 feature cells audited under each rule

Hindsight enters through two doors

What this shows. Strict as-of uses only information known at the time. Fixed mask keeps the original release availability but imports 5,464 later revisions. Unrestricted latest values also introduce 863 future releases. The two counts can overlap; do not add them.

View exact values & source
Information ruleFuture-release cellsFuture-revision cells
As-of00
Fixed mask05464
Naive revised8636311

Source: Macroeconomics/data/generated/official_pilot/feature_leakage_audit.parquet

Download figure (SVG)

All six GDP models change rank

Same availability; different data revisions · 8 final holdout forecasts · horizon 0

All six GDP models change rank

What this shows. In the final GDP holdout, Elastic Net moves from fourth to first when later revisions replace the original values. AR(1) moves from first to second. Both predictor values and scoring targets are revised, while release availability stays fixed. Eight forecasts (July 2024–April 2026) are too few to establish general model superiority.

View exact values & source
ModelStrict as-of rankRevised fixed-mask rank
AR(1)12
Bridge65
Elastic Net41
Gradient boost34
Historical mean23
No change56

Source: Macroeconomics/data/generated/official_pilot/model_stability.parquet

Download figure (SVG)
Research notes

Design, sources & limitations

Research contribution

I built the archive-to-backtest workflow: official release histories, information sets reconstructed at each forecast date, comparable predictions, and explicit checks for future information.

The strict as-of rule has no future releases or later revisions. Fixed mask preserves release availability but uses 5,464 later-revised cells. The naive latest-data rule adds 863 future-release cells as well. These are three distinct information rules, not three valid real-time forecasts.

Research question and relevance

Question: How do release revisions and the information available at each vintage affect nowcast and forecast quality?

Real-time evaluation matters because today's revised macro series contain information unavailable to a forecaster in the past. The project makes availability dates, vintage labels, transformations, and source limitations explicit.

Original paper or report: This is a project-originated forecasting engine. Any future paper replication must add the exact citation, original sample, specification, and benchmark result before claiming equivalence.

Data sources and sample

Adapters support FRED and ALFRED plus guarded BLS and BEA current-data APIs. Synthetic fixtures allow offline validation. Current BLS and BEA API rows are labeled latest_revised; they are not presented as first-release historical observations.

SourceCoverage and provenance
FRED / ALFREDSeries-specific sample and vintage metadata.
BLS / BEACurrent revised public API data under explicit authorization.
FixturesSynthetic offline samples for schema and pipeline tests.

Capabilities demonstrated

  1. Record source retrieval time and release availability.
  2. Normalize units, frequency, adjustment, and transformations.
  3. Construct vintage-consistent information sets.
  4. Backtest models without revised-data leakage.
  5. Publish forecast outputs with source and revision manifests.

Robustness checks cover alternative windows, model classes, transformations, missing-value behavior, retry paths, and unsupported provenance claims.

Validation evidence

Validated componentEvidenceInterpretation
Official archive ingest626,304 rows across 21 seriesInputs are hashed, provenance-stamped, and aligned to their documented release timing.
Real-data forecast run15,264 two-horizon forecastsThe expanding-window pilot compares models on consistent information sets.
Vintage timing audit0 future-release and future-revision cells under strict as-ofFixed mask isolates revised-value hindsight; naive latest data adds release-timing leakage. Neither is a valid historical real-time information set.

This is an official-data pilot and research system, not a claim to reproduce a specific paper or production forecast.

Reproduction and citation

Run python -m macro_nowcast.pipeline using the documented configuration. Cite the portfolio release, Dataset revision, project, and original agency sources.

Evidence boundary

The official pilot distinguishes historical archives from current revised APIs and records the vintage policy, sample, configuration, metrics, and checksums. It is not presented as a paper replication or a production forecast.